Plan a training run

From reference math to target-speed kernels.

SF Tensor automatically explores hardware-specific implementations of your workload, finds the fastest viable candidates and verifies that optimized code still computes what you asked for.

Bring us a workloadSee how it works
Kernel optimizationAutomatic

Reference

Same calculation

Search + model

Rank candidates

Optimized kernel

Target specific

Search

Equivalent programs

Measure

Target hardware

Prove

Contract passed

Input

A mathematical specification, not a hardware-specific implementation.

Search

Architecture-aware candidates generated and ranked automatically.

Output

A measured kernel with correctness evidence attached.

Why kernels stay hard

The fastest kernel is specific to the chip. Your model shouldn’t have to be.

High-performance kernels are shaped by a target’s instructions, memory system and execution topology. A portable language can translate syntax, but it cannot make one kernel design optimal for every architecture.

Our Kernel Optimizer treats the calculation as the constant and the implementation as a search problem. That lets the system create a different kernel for each target without asking researchers to rewrite the model.

Memory hierarchy

Model access patterns, placement, cache behavior and the cost of moving data through the chip.

Execution topology

Account for instructions, lanes, warps, cores and synchronization on the actual target.

Measured hardware profile

Use automatic microbenchmarks to calibrate the parameters that matter for ranking candidates.

How it works

Search the implementations. Preserve the calculation.

Traditional compiler passes improve a program while keeping its structure mostly intact. We search for a new program that performs the same calculation efficiently on a specific target.

01

Specify the calculation

Start with the mathematical operation the workload needs, expressed as a clear reference rather than a hand-tuned implementation.

02

Explore equivalent programs

Generate candidates across algebraic rewrites, fusion, tiling, loop order, memory placement and execution schedule.

03

Prune, model and measure

Eliminate poor candidates early, rank the survivors against the target architecture and benchmark the most promising kernels.

04

Prove before production

Compare the optimized kernel with its specification, reject semantic errors and return concrete counterexamples when a proof fails.

Equivalent program spaceSearch 02

Specification

attention(q, k, v)

Fuse QK
Stage QKV
Async copy

64×64

128×64

128×128

64×128

PruneRankMeasure
Beyond autotuning

Don’t benchmark billions of possibilities. Make the search intelligent.

A fully expanded search space for a modern kernel can contain billions of valid-looking programs. Compiling and benchmarking every one would turn optimization into its own compute problem.

Prune obvious losses

Estimate dynamic cost and discard candidates with untenable allocation, movement or execution costs.

Rank what remains

Model the target’s memory, cache and inter-core topology to focus measurement on promising designs.

Learn from the hardware

Use real measurements and automatically generated microbenchmarks to calibrate each target profile.

Correctness built in

Optimization should end with a proof obligation.

Fast code is only useful if it is still the right code. Our proof system symbolically compares a generated kernel with its reference over the input space represented by the contract, rather than trusting a finite sample of tests.

Equivalence proofContract 03

Reference

≠ ?

Candidate

Symbolic solver
ALL INPUTS

No counterexample exists

Algebraically equivalent

UNSAT
Workload-specific

Optimize the operations that make your model yours.

The target is not a generic benchmark. It is the operator, shape, precision and hardware combination that runs inside the production workload.

Sequence workloads

Attention

Specialized forward and backward kernels for sequence length, head dimension, masking and precision.

Core model compute

Linear algebra

Matrix operations tuned around the shapes, data types and layouts that appear in the real model.

Bandwidth-sensitive

Normalization & reductions

Fuse memory-bound operations and choose schedules that reduce traffic and synchronization overhead.

Custom research

Novel operators

Give new architectures a path from readable reference code to hardware-specific implementation.

Bring us the kernel slowing your model down.

We’ll look at the workload, target hardware and performance constraint with you, then map the shortest path to a production-ready implementation.

Talk to an engineer

Train the models only you can build. One stack for enterprise post-training and frontier pre-training.

All Systems Operational

© 2026 San Francisco Tensor Company

SolutionsHomeEnterprisesAI LabsTensor CloudKernel OptimizerEmma LangTalk to an engineer
Company
Blog
Manifesto
CareersWe Are Hiring!
Engineering
SF Tensor Stamp